Expert Judgement and Big Data in Reliability
Role of Expert Judgement in Analysing Large
Complex Reliability Data Sets
Expert Judgement and Big Data in Reliability
1
Background to the problem
2
Empirical analysis of observational condition test data
3
Engineering judgement reasoning of the problem example
4
Building a predictive model for inference about failure
Expert Judgement and Big Data in Reliability Background to the problem
Motivating engineering problem
Predict the time to failure of vehicles given information about engine condition
Our problem of interest today
Mining observational condition monitoring data to understand patterns and correlation structures
Compare empirical insights with engineering judgement of causal relationships
Use relevant to inform predictive model development
Condition monitoring observational data
Relate to fleet of like vehicles with irregular usage profile Engine oil samples taken at regular intervals
Expert Judgement and Big Data in Reliability Background to the problem
Expert Judgement and Big Data in Reliability Background to the problem
Data initially in 3 spreadsheets by vehicle type.
Dimensions are (437
×
41), (286
×
41) and (70
×
41).
The covariates are the same in all sheets but there are many
missing values and entire missing covariates in some sheets.
The data were messy and required initial cleansing.
Specific variables:
Distances travelled since oil change for each vehicle at each observation: km and motor hours.
Number of indicator particles in oil: cutting, sliding, fatigue, total, “large”.
Content of specific elements in oil: chromium, lead, copper, nickel, etc.
Other measures: kinematic viscosity, flash point, total alkilinity number.
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
km.tot km.oil mh.tot mh.oil viscosity flash alkalinity
fuel water glycol oxid nitrate sulphate antioxid
soot ferrum chromium lead copper tin aluminium
nickel silver silicon borium natrium magnes calcium
barium phosphor zinc molybdenum titanium vanadium part.tot
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
0 20 40 60 −1.0 −0.5 0.0 0.5 1.0 Correlation count
Variable 1
Variable 2
Correlation
Total km
Total motor hours
0.90
Fuel content
Flash point
-0.90
Zinc content
Calcium content
0.88
Chromium content
Ferrum content
0.86
Sulphate content
Nitrate content
0.71
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
km.tot km.oil mh.tot mh.oil viscosity flash alkalinity fuel oxid nitrate sulphate antioxid soot ferrum chromium lead copper aluminium nickel silver silicon borium natrium magnes calcium barium phosphor zinc molybdenum vanadium part.tot cut.part slid.part fat.part large.part
km.tot km.oil mh.tot mh.oil viscosity flash alkalinity fuel oxid nitr
ate
sulphate antio
xid
soot ferr
um
chromium lead copper aluminium nick
el silv er silicon bor ium natr ium
magnes calcium bar
ium
phosphor zinc molybden
um v anadium par t.tot cut.par t slid.par t fat.par t large .par t −1.0 −0.5 0.0 0.5 1.0
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
Two algorithms to determine Bayesian network structure from
data: Growth-Shrink algorithm and Hill-Climbing algorithm.
The Growth-Shrink algorithm uses Markov Blankets to detect
the structure of the network.
Definition
A Markov Blanket
BL
(
X
) for some variable
X
∈
U
is a set of
variables which, for any further variables
Y
∈
U
−
BL
(
X
)
−
X
, has
the property that
X
,
Y
are independent given
BL
(
X
).
The Hill-Climbing algorithm works as follows:
1 Start from an empty, full or random network.
2 Attempting every possible single-edge addition, removal, or
reversal.
3 Choose the move which increases the “score” by the most.
4 The process stops when there is no single-edge change that
increases the “score”.
The score in question can either be “log-likelihood”, “AIC”, or
“BIC”.
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
Grow−shrink nr km.tot mh.tot viscosity flash alkalinity fuel oxid nitrate sulphate antioxid soot ferrum chromium lead copper aluminium nickel silver silicon borium natrium magnes calcium barium phosphor zinc molybdenum vanadium part.tot cut.part slid.part fat.part large.part Hill−climbing nr km.tot mh.tot viscosity flash alkalinity fuel oxid nitrate sulphate antioxid soot ferrum chromium lead copper aluminium nickel silver silicon borium natrium magnes calcium barium phosphor zinc molybdenum vanadium part.tot cut.part slid.part fat.part large.part
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
0 5 10 15 20 25 30 35 0 2 4 6 8 Scree Plot Dimension Eigen v alue ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● Grow−shrink
f1
f2
f3
f4
f5
f6
Factor descriptions:
Factor 1: chemical elements that reflect the primary materials of key dynamic (as in moving) engine parts.
Factor 2: operational exposure over both degradation paths of oil and engine condition.
Factor 3: operational exposure over degradation in confounded degradation path of oil and engine condition.
Factor 4: measures of oil quality.
Factor 5: chemical elements arising from other static. Factor 6: measures of oil morphology.
Expert Judgement and Big Data in Reliability Empirical analysis of observational condition test data
km.tot km.oil mh.tot mh.oil viscosity flash alkalinity fuel oxid nitrate sulphate antioxid soot ferrum chromium lead copper aluminium nickel silver silicon borium natrium magnes calcium barium phosphor zinc molybdenum vanadium part.tot cut.part slid.part fat.part large.part Grow−shrink algorithm
Expert Judgement and Big Data in Reliability
Engineering judgement reasoning of the problem example
Expert
Highly qualified (PhD) mechanical engineer with operational experience of vehicles
Process
Motivate and condition to provide engineering reasoning from cause to effect to establish relationships between engine condition and observables
Structure around key questions to build causal maps of reasoning using why and how laddering
E.g. if condition of engine is X then observation on test is Y (HOW manifest)
E.g. if observe Y on test then state of engine condition is X (WHY might see)
Consistency checking the causal reasoning
Real-time reasoning verification by dual questions and preliminary quantification
Multiple (3) stages of cross-questioning as facilitator learns
Limitations
Only one expert with knowledge domain of the operation, test and maintenance of vehicles
Expert Judgement and Big Data in Reliability
Expert Judgement and Big Data in Reliability Building a predictive model for inference about failure
We can model the degredation of the oil morphology through
the numbers of Cut, Slide and Fatigue particles.
Y
t,j,k=
t
X
i=1
X
i,j,k,
for vehicle
j
, measure
k
and exposure time point
t
.
Define
Zt
,j,k=
log
Yt
,j,kY
t−1,j,kd
j,k(
t
)
−
d
j,k(
t
)
,
where
dj
,k(
t
) is the exposure of vehicle
j
at time point
t
.
Assume parametric forms
Z
t,j,k∼
G
(1
, µ
j,k)
,
µ
j,k∼
IG
(
α
k, β
k)
.
Expert Judgement and Big Data in Reliability Building a predictive model for inference about failure
Expert Judgement and Big Data in Reliability Summary and next steps
Our problem
is a typical general asset management problem although presented for specific data
Our observational data
are real and typical of condition monitoring data structures although our frequency of observation is less than some other assets with real-time applications (e.g. wind farms)
Our judgemental data process
challenges the engineer to think through causal reasoning from true condition state to observable measures of condition
Our analysis
reveals insights into life and condition of vehicles that align with engineering expectations about first order dependencies highlights more subtle dependency structures that should provide information about engine condition
exposes the usual limitations of data mining without contextual knowledge e.g. sample number in blind analysis!
Our next step is
To use the insights gained to develop a predictive model grounded in the observations and the judgements
Expert Judgement and Big Data in Reliability Summary and next steps